Content
90%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A dense, well-orchestrated skill body: executable commands with exact flags, an unambiguous identity-state workflow, and a genuinely progressive loading design that keeps the overview lean. The main gaps are validation checkpoints delegated to scene files and the absence of the referenced scenes/ files (and cross-skill lark-drive references) from the bundle, which weakens the otherwise excellent navigation layer.
Suggestions
Ship the six scenes/*.md handbooks (and the two ../lark-drive reference targets) with the bundle, or inline a minimal fallback flow for each scene in SKILL.md, since the body routes most non-quick-action tasks to them.
Surface one or two explicit verification checkpoints for write operations in the body itself (e.g. confirm meeting_id and participant list before `vc +meeting-end` or `drive +member-add`) rather than relying solely on the scene handbooks.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and assumes Claude's competence: no explanation of what Feishu, meetings, or calendars are; the domain-model diagram and table carry only non-obvious internal relationships (note_display_type branches, meeting_note vs AI 纪要, verbatim normal/unified). Every section adds CLI-specific or routing knowledge Claude could not infer, matching 'every token earns its place'. | 5 / 5 |
Actionability | 快速行动 gives copy-paste-ready commands with exact flags ("lark-cli vc +meeting-list-active --as bot --user-id <open_id>", "lark-cli vc +meeting-events --as <source_identity> --meeting-id <meeting_id> --page-all --format pretty"), and the command table maps all 26 commands to a concrete lookup path (a specific references/*.md file or an exact `lark-cli ... --help` invocation). The identity procedure is fully concrete decision logic, not vague direction. | 5 / 5 |
Workflow Clarity | Multi-step processes are clearly sequenced: numbered identity-selection with explicit stop conditions ("不支持:说明限制并停止"), a 3-step conditional 渐进加载规则, a multi-meeting disambiguation rule, and an auth-error feedback loop ("遇到未认证…读取该 Skill,修复后重试"). Falls short of 5 because explicit validation checkpoints for write operations (end meeting, permission changes, minutes edits) are delegated to scene-handbook files rather than shown, leaving the in-body checkpoints mostly for read and routing flows. | 4 / 5 |
Progressive Disclosure | Exemplary one-level-deep design: quick action inline, six clearly-signaled scene handbooks, a per-command reference table, and an explicit progressive-loading rule section; all 22 references/*.md files referenced in the command table exist in the bundle and contain substantive content (verified). Held at 4 rather than 5 because the six scenes/*.md handbooks — the primary flow layer the body routes most tasks to — plus two ../lark-drive references are not present in the bundle, so the prescribed navigation path dead-ends for non-quick-action tasks (the --help fallback in loading rule 3 partially mitigates this). | 4 / 5 |
Total | 18 / 20 Passed |